Lane Detection and Traffic Sign Detection using Deep Learning and Computer Vision for Autonomous Driving Research Using CARLA Simulator

Abstract

Lane identification and traffic sign detection is the most challenging and promising problem for self-driving or autonomous vehicles with unintentional lane departure and ignorance of traffic signs being major contributing factors to motor vehicle collisions around the world. To tackle this problem the proposed work aims to detect both lane and traffic signs for autonomous vehicles. This article proposes semantic segmentation and object detection model for implementing Advanced Driver Assistance System (ADAS) applications. The applications are implemented using a variant of Convolutional Neural Networks (CNN) deep learning model such as SegNet and You Only Look Once (YOLO) algorithm. Due to dynamic and adverse environment conditions, devising and testing a system which yields effective performance in all urban driving scenarios is challenging. Hence the environment is set up virtually with the help of the Car Learning to Act (CARLA) simulator. With aid of developed models, lane and traffic signs were successfully detected and tested under various constraints. Obtained results are evaluated with various performance metrics.  Models are deployed for separate created datasets. The semantic segmentation model developed for lane detection using SegNet gives a mean average precision (mAP) of 93.33%, an overall accuracy of 94.80%, F-score of 93.42% and a minimal error rate of 5.20%. Model developed for Traffic sign detection, a mean average precision of 93.67%, an accuracy of 95.56%, recall of 92.67%, F-score of 93.16% and error rate of 4.44% have been achieved.